Author: Richard A. Fowell (fowell@netcom.com)
Date: 07:15:23 09/15/99
Go up one level in this thread
The consumer version was mentioned in this article, which I have summarized
here.
The magazine in question would not be happy if they found Hsu was using them
to propagate a hoax. Or are you calling me a liar? Granted, many things could
get in the way of Hsu's plans.
Richard A. Fowell (fowell@netcom.com)
http://dmoz.org/Games/Board_Games/Chess/Software/Macintosh/
====== The below is my abstract after reading the article ========
Consumer Version of Deep Blue Planned ...
There is an 11 page article by Deep Blue's hardware architect,
Feng-hsiung Hsu, in IEEE Micro, Mar/Apr 1999, pp. 70-80, discussing
the chips used to defeat Kasparov, and future plans. Hsu is listed
as the sole author, so the article is presumably authoritative.
The title is "IBM's Deep Blue Chess Grandmaster Chips". The journal
should be available in most university or large corporate libraries.
I'll summarize some of the highlights here. [N] should be read as,
"On page N, Hsu says ..." Parenthesized comments are mine.
1) Deep Blue for Consumers
[81] Hsu is forming an independent start-up to create a consumer chess
chip for desktop machines. A desktop machine using this chip (many, I assume
- RAF) could "defeat the World Champion in a formal match as early as the year
2000." [71] 30 million positions/sec is possible with a 0.35 micron process
and a new design. [71] The 1997 chips searched 2-2.5 million. [71] "Such a
chip might make it possible to defeat the World Chess Champion with a desktop
personal computer or even a laptop".
2) Deep Blue Jr.
[80] Deep Blue Jr. uses multiple chips on a single workstation.
[71] Each RS/6000 controlled up to 16 chess chips.
( This suggests that Deep Blue Jr. is a 16-chip configuration.-RAF )
3) Deep Blue vs the micros - the controversial "10 game match"
[72] Kasparov's opponent used 480 chips. [80] A single chip, derated to
10-20% of normal speed, scored 10-0 against "Two of the top commercial
programs, running on the Pentium Pro PC." ( The significance of this has been
hotly debated, given the variance of small matches. -RAF ) [80] This
indicates that a single, slowed-down chip is 200 points stronger than the
commercial programs, at roughly a 95% confidence level.
[80] In 30 later games against "the single chip version or Deep Blue Jr."
(fraction of games by each, and speed derating, if any, not stated - RAF)
the commercial programs scored 2-38. ( Time control unstated, however:
[71] the important measure of strength is under regulation time control
- "no faster than 3 minutes a move". - RAF)
4) Deep Blue vs. humans
[81] In training games against Grandmasters with average FIDE ratings in
the "high 2500s", Deep Blue Jr. scored at a 3-1 ratio, for a performance
rating of over 2700. [72] Kasparov's opponent used many more chips - 480.
( Time control unstated, however: [71] the important measure of strength is
under regulation time control - "no faster than 3 minutes a move". - RAF)
5) Deep Blue search depth vs. Kasparov
[72] In 1997, the non-extended ply depth was about 12 plies, with about 40
plies along extended lines. In an exemplary 12 ply search, the master
RS/6000 would search 4 plies, generating about 1000 positions. These would
be searched in turn by the 30 RS/6000s to four more plies, for about
1,000,000 leaf positions. The chips would then search all of these
1,000,000 leaves to four plies, including quiescence search. [80]
A search can be initiated from the position on a chip, usually for
four or five plies.[80] Search control uses a minimum window alpha-beta search.
6) Deep Blue evaluation function
[72] On average, each position evaluation consumes 40,000 general-purpose
instructions. [75] The chip includes both a "fast evalation", followed,
if necessary, by a "slow evaluation". (This seems to be a "lazy evaluation" -
-RAF) [78] Fast evaluation takes one cycle, slow evaluation 11 cycles.
[79] At typical search depths, only about 15% of positions require the
slow evaluation.
[78] Fast evaluation circuitry includes:
- Piece Placement circuit
- Pawn run circuit
Passed pawns/pawn pairs unstoppable by King or Rook
- Endgame circuit
- [79] KPvK, KRvKP, KQvKP, KRPvKP, plus knowledge of very drawish
simple endings and some simple fortress draws.
- phase control
- adjusts for certain piece combinations
- adjust other weights for
- king safety relevance
- bad pawn structure relevance
- passed pawn relevance
[79] Slow evaluation circuitry (close to half of the chip core)
- Includes: trapped pieces, development, restraint, color complex,
blockade, square control, king safety, pawn structure, outposts,
passed pawns, pawn majority, rook on the 7th, color complex,
x-rays, pins, ray control, etc.
The evaluation is very complicated - for example, king safety includes:
- types of attacking pieces
- soundness of king shelter
- presence of attacking pawns
- around the king:
- color complex
- square control
- ray control
7) Deep Blue chip hardware parameters for the 1997 version
[71] Searches 2-2.5 million chess positions/second
[71-72] Eight chips per Micro Channel card, two cards per workstation.
[75] Die photo is Figure 2
[75] 0.6 micron, 5-V CMOS, 3-metal layer
[75] 40-50 nanosecond cycle time
[75] About 1 watt power consumption
[76] About 1.5 million transistors
[76] About 1.4cm x 1.4 cm die size
[80] Appears as a 32-bit device with a 17 bit address space
[71] Positional weights software-adjustable
Chip partitioning, excluding ROM and RAM:
[78] 52,400 gates - move generator
[75] ??,??? gates - smart-move stack
[72] ??,??? gates - regular move stack
[80] 20,000 gates - repetition detector
[78] 66,000 gates - evaluation function (which has a lot of ROM and RAM)
[72] ??,??? gates - fast evaluation
[72] ??,??? gates - slow evaluation
[75] ??,??? gates - alpha-beta search control
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.